Abstract:Back-propagation neural-networks (BP-NN) and the support vector machine (SVM) are the two mainstream methods for classification of seismic wave signals of earthquakes and explosion events used in this research. The two methods achieved accurate and effective results. However, when training the BP-NN, it is inevitable that it can be easily trapped in a local optimum; in addition, the optimal numbers of hidden layers and numbers of nodes in each layer are heavily dependent on the distribution configuration of the training samples data, and cannot be consistently determined in advance. Furthermore, when training the SVM, there is a shortage of effective means to select suitable kernel function(s); hence, the ordinary SVM cannot be easily extended to multiclass problems. Aiming at the classification of seismic wave signals of earthquakes and explosion events, this paper investigates and compares the BP-NN and the SVM, along with the BP-Adaboost ensemble learning method. Using the dataset of seismic wave signals of earthquakes and explosion events in the experiments of this paper, the classification results show that the BP-Adaboost method can achieve the overall correct recognition rate of not less than 98%, with excellent generalization ability. Compared with BP-NN and SVM, the two main traditional classification methods, it has been shown that the BP-Adaboost method is more robust for different dataset partitions and corresponding classification, which implies more robust generalizability and better classification of seismic wave signals of earthquakes and explosion events. In the meanwhile, the theory of the Adaboost method is applied to explain the reasons for the better classification results and the generalizability of the BP-Adaboost method.